AMD Ryzen 9 3900XT review
AMD Ryzen 7 3800XT review
AMD Ryzen 5 3600XT review
Crucial P2 500 GB NVMe M2 SSD review
ASUS STRIX B550-I Gaming review
Kioxia Exceria 1TB M.2 NVMe SSD review
Samsung 870 QVO 2TB SSD review
Guru3D Rig of the Month - June 2020
ASUS ROG Strix Z490-E Gaming review
MSI MEG Z490 Godlike review
AMD Could Do DLSS Alternative with Radeon VII through DirectML API
There is an interesting development, as you know the GeForce RTX graphics cards have tensor cores, dedicated cores for optimizing and accelerating anything AI and Deep Learning. A spawn from that technology for gaming is DLSS, which NVIDIA is marketing strongly. With the October 2018 update for Windows 10, Microsoft has released the DirectML API for DirectX 12.
ML stands for Machine Learning - and makes it possible to run trained neural networks on any DX12 graphics card. Basically, the Deep Learning optimization or algorithm if you will, becomes a shader that can be run over the traditional shader engine, without a need for tensor cores. In an interview with AMD they mention that the team is working on this and it seems, Radeon VII seems very well suited for the new API. Japanese website 4gamer.net spoke with AMD marketing manager Adam Kozak, AMD is currently testing DirectML with Radeon VII and was positively impressed by the results, and that is exciting news for AMD offering them an AI/DL alternative.
While AMD is testing this on Radeon VII, logic obviously dictates that it would work well on Radeon RX Vega 64 and Radeon RX Vega 56 as well. This allows, for example, an alternative implementation to Nvidia's DLSS.
Only 1+1=2
Of course, should this become an actual supported thing, then it can't be addressed AMD alone, this question remains: will game developers actually implement, experiment and integrate support for DX ML into games?
It also has to be said, it works reversed, DirectML could also make use of Nvidia's Tensor cores, certainly giving them an advantage. As the Radeon card would see a performance hit, whereas the RTX cards can simply offload the algorithm towards its Tensor cores. Time will tell, but this certainly is an interesting development as long as your graphics card is powerful enough, of course.
AMD Could take Back 30% of the Processor Market - 09/26/2018 07:55 AM
Good times for AMD. Intel is under a lot of scrutinies lately, scandals with their top-tier staff, issues with 14nm production, delays at 10nm and vulnerabilities are only a few of them. Meanwhile...
AMD Could Potentially Get 19B investment - 02/11/2015 10:48 AM
More news from AMD today as it seems the company might get a nice cash injection. Loongson Technology, a CPU joint venture between Beijing-based chip designer BLX IC Design Corp, the Chinese Academy ...
AMD could be restructuring once more - 10/06/2014 10:04 AM
BSN reported earlier that AMD might be preparing one more round of restructuring, a pretty significant one as well. The reorganization would be announced later this month. Sources claim the restruct...
AMD could ship 28nm GPUs in December - 10/18/2011 10:45 AM
We can confirm this rumor as AMD has been 'carefully' talking about it. AMD might still be planning to introduce some 28nm GPUs in the second week of December. It's expected that these chips will be ...
AMD could ship 8 million Llano APUs this year - 08/05/2011 10:22 AM
AMD is doing well with Llano alright, Sources at motherboard makers told DigiTimes that AMD shipped about 1 million Llano APUs in June and 1.3-1.5 million units in July. Total shipments for 2011 are e...
-Tj-
Senior Member
Posts: 16845
Joined: 2012-05-18
Senior Member
Posts: 16845
Joined: 2012-05-18
#5628899 Posted on: 01/17/2019 06:53 PM
I rather have directML then dlss. At least when I saw that car reconstruction picture.
The biggest reason is quality, unless you use 2x dlss to get over that upsampling , but then it's kind of a moot point - no perf boosts..
I saw really detailed review about dlss @ ffxv and to be honest it looked crap 90% of the time.
The worst part was fence lines shimmering and some smeared pixels with loss of texture detail and even object detail in the distance.
I rather have directML then dlss. At least when I saw that car reconstruction picture.
The biggest reason is quality, unless you use 2x dlss to get over that upsampling , but then it's kind of a moot point - no perf boosts..
I saw really detailed review about dlss @ ffxv and to be honest it looked crap 90% of the time.
The worst part was fence lines shimmering and some smeared pixels with loss of texture detail and even object detail in the distance.
BlackZero
Senior Member
Posts: 8880
Joined: 2007-06-17
Senior Member
Posts: 8880
Joined: 2007-06-17
#5628902 Posted on: 01/17/2019 07:06 PM
Clearly, all this adds latency, but we were discussing older GPUs that probably aren't achieving 60 FPS in the first place.
If it's worth developer time or not, that I suppose would depend on if any real benefit can be attained.
It's more about transfer time (plus eventual decoding) then computation time, especially from GPU to CPU, readback operation can become easily a bottleneck since they break rendering pipeline. Also, abuse of CPU to GPU upload can become a problem too, especially on discrete GPUs.
Clearly, all this adds latency, but we were discussing older GPUs that probably aren't achieving 60 FPS in the first place.
If it's worth developer time or not, that I suppose would depend on if any real benefit can be attained.
Denial
Senior Member
Posts: 12784
Joined: 2004-05-16
Senior Member
Posts: 12784
Joined: 2004-05-16
#5628906 Posted on: 01/17/2019 07:18 PM
Isn't this one of reasons AMD is working on HBCC and IF to be able share data with minimal latency?
Yes.
Plus I am pretty sure data transfer and operations between CPU and GPU takes microseconds and not milliseconds unless said operation take hundreds of clock cycles.
Especially on Zen where through IF different resources could have direct access without need to wait for CPU cores to do all actions.
The latency would depend on the size of the data but it's not really relevant. In this case Microsoft found GPU processing on DirectML with metacommands on to be 275x faster then running it on the CPU.
http://on-demand.gputechconf.com/siggraph/2018/video/sig1814-2-adrian-tsai-gpu-inferencing-directml-and-directx-12.html - @24 minutes into presentation - the entire presentation is good though and covers a lot stuff being said here.
Point is even if the latency is only 100-200us to transfer the CPU, the GPU could have performed whatever operation that was sent to CPU multiple times over again.The more data you send the longer the time to get it back. It's simply never worth sending it there - especially with the order of magnitude in performance.
I am not saying that CPU must be able to do it (performance) but latency between CPU, GPU, memory and cache should not be big deal unlike operations itself.
Question here is about controlling stuff, but this is something that needs to be solved anyway to be able use chiplets in GPU and still be visible as one GPU unlike current SLI/CF and its something AMD is likely working on and Nvidia probably too.
https://hps.ece.utexas.edu/people/ebrahimi/pub/milic-micro17.pdf
They both are working on it but it requires massive amounts of bandwidth, changes to the scheduling, etc and even then it's still not scaling perfectly in terms of performance.
BTW, in big data and ERP multi-node systems you have server to server (each different physical frame) data latency in range of 100+ microseconds and that's for two systems that have to talk through network where network latency is bottleneck.
You have a source for 100 microseconds? Typically the latency between two multnode systems ~350-400us for the network alone - but admittedly it's been a while since I worked on anything like this (2011/12 @ RIT).
Isn't this one of reasons AMD is working on HBCC and IF to be able share data with minimal latency?
Yes.
Plus I am pretty sure data transfer and operations between CPU and GPU takes microseconds and not milliseconds unless said operation take hundreds of clock cycles.
Especially on Zen where through IF different resources could have direct access without need to wait for CPU cores to do all actions.
The latency would depend on the size of the data but it's not really relevant. In this case Microsoft found GPU processing on DirectML with metacommands on to be 275x faster then running it on the CPU.
http://on-demand.gputechconf.com/siggraph/2018/video/sig1814-2-adrian-tsai-gpu-inferencing-directml-and-directx-12.html - @24 minutes into presentation - the entire presentation is good though and covers a lot stuff being said here.
Point is even if the latency is only 100-200us to transfer the CPU, the GPU could have performed whatever operation that was sent to CPU multiple times over again.The more data you send the longer the time to get it back. It's simply never worth sending it there - especially with the order of magnitude in performance.
I am not saying that CPU must be able to do it (performance) but latency between CPU, GPU, memory and cache should not be big deal unlike operations itself.
Question here is about controlling stuff, but this is something that needs to be solved anyway to be able use chiplets in GPU and still be visible as one GPU unlike current SLI/CF and its something AMD is likely working on and Nvidia probably too.
https://hps.ece.utexas.edu/people/ebrahimi/pub/milic-micro17.pdf
They both are working on it but it requires massive amounts of bandwidth, changes to the scheduling, etc and even then it's still not scaling perfectly in terms of performance.
BTW, in big data and ERP multi-node systems you have server to server (each different physical frame) data latency in range of 100+ microseconds and that's for two systems that have to talk through network where network latency is bottleneck.
You have a source for 100 microseconds? Typically the latency between two multnode systems ~350-400us for the network alone - but admittedly it's been a while since I worked on anything like this (2011/12 @ RIT).
holler
Senior Member
Posts: 221
Joined: 2003-07-07
Senior Member
Posts: 221
Joined: 2003-07-07
#5628911 Posted on: 01/17/2019 07:27 PM
i could really care less about DLSS, its image quality improvements are of questionable value. just because you can do it, doesn't mean you should...
i could really care less about DLSS, its image quality improvements are of questionable value. just because you can do it, doesn't mean you should...
Click here to post a comment for this news story on the message forum.
Senior Member
Posts: 1804
Joined: 2015-06-11
Because sending a lot of 0s and 1s takes up huge amounts of CPU time and bandwidth?
Anyway, if they could make it happen, it could be useful.It's more about transfer time (plus eventual decoding) then computation time, especially from GPU to CPU, readback operation can become easily a bottleneck since they break rendering pipeline. Also, abuse of CPU to GPU upload can become a problem too, especially on discrete GPUs.